fix(core): lazy loading transformers, numpy, simsimd

[Won-Kyu Park (wkpark)]

fix(core): lazy loading transformers, numpy, simsimd

Summary

Optimized the initial import speed of langchain_core by converting top-level imports of transformers, numpy, and simsimd into lazy, local imports.

Details

Confirmed with python -X importtime that these dependencies are no longer loaded on initial import, and verified functionality through make format/lint and est.

Disclaimer

I used the Gemini CLI to help with the specific implementation details and to ensure compliance with linting and type checking and drafting this PR description.

before
without transformers / torch env.

$ time uv run python -c "from langchain_core.prompts import BasePromptTemplate"
real    0m0.600s
user    0m0.509s
sys     0m0.074s

with transformers / torch installed env.

$ uv pip install transformers
...
$ uv pip install numpy
...

$ time uv run python -c "from langchain_core.prompts import BasePromptTemplate"
real    0m9.664s
user    0m4.268s
sys     0m0.752s

after

$ time uv run python -c "from langchain_core.prompts import BasePromptTemplate"
real    0m1.916s
user    0m0.802s
sys     0m0.132s

[codspeed-hq]

Merging this PR will improve performance by 20.01%

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚡ 1 improved benchmark
✅ 12 untouched benchmarks
⏩ 23 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	WallTime	test_import_time[InMemoryVectorStore]	651 ms	542.5 ms	+20.01%

---

_{Comparing wkpark:lazy_load_numpy_transformers (b5f28ac) with master (e939c96)}

Footnotes

1. 23 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩